fix: update to python 3.12.

dockerfiles/Dockerfile.ray-worker-strixhalo

@@ -6,14 +6,15 @@
 #   docker build -t git.daviestechlabs.io/daviestechlabs/ray-worker-strixhalo:latest \
 #     -f dockerfiles/Dockerfile.ray-worker-strixhalo .
 #
-# Multi-stage build: Extract ROCm 7.1 from vendor image, use Ray base for Python 3.11
+# Multi-stage build: Extract ROCm 7.1 from vendor image, use Ray base for Python 3.12
 # Note: Uses TheRock gfx110X wheels due to ROCm/ROCm#5853 segfault issue
+# Note: Python 3.12 required — vLLM ROCm wheel (wheels.vllm.ai/rocm) is cp312 only
 
 # Stage 1: ROCm 7.1 libraries from AMD vendor image
 FROM docker.io/rocm/pytorch:rocm7.1_ubuntu24.04_py3.12_pytorch_release_2.9.1 AS rocm-source
 
 # Stage 2: Production image
-FROM docker.io/rayproject/ray:2.53.0-py311 AS production
+FROM docker.io/rayproject/ray:2.53.0-py312 AS production
 
 # OCI Image Spec labels
 LABEL org.opencontainers.image.title="Ray Worker - AMD Strix Halo"
@@ -59,11 +60,16 @@ COPY --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/uv
 
 USER ray
 
-# Install vLLM and inference dependencies first (without torch)
-# vLLM will try to install CUDA torch as dependency, we exclude it here
+# Install vLLM ROCm build and inference dependencies.
+# The vLLM ROCm wheel from wheels.vllm.ai includes HIP-compiled C-extensions
+# (vllm._C, vllm._rocm_C) that are ABI-compatible with ROCm PyTorch.
+# PyPI vLLM is CUDA-only and crashes with: libcudart.so.12 not found.
+# uv gives --extra-index-url higher priority than PyPI, so the ROCm wheel
+# is selected over the CUDA wheel.
 RUN --mount=type=cache,target=/home/ray/.cache/uv,uid=1000,gid=1000 \
     uv pip install --system \
-        'vllm>=0.5.0' \
+        --extra-index-url https://wheels.vllm.ai/rocm/ \
+        vllm \
         'transformers>=4.35.0,<5.0' \
         'accelerate>=0.25.0,<1.0' \
         'sentence-transformers>=2.3.0,<3.0' \
@@ -72,8 +78,9 @@ RUN --mount=type=cache,target=/home/ray/.cache/uv,uid=1000,gid=1000 \
 
 # WORKAROUND: ROCm/ROCm#5853 - Standard PyTorch ROCm wheels cause segfault
 # in libhsa-runtime64.so during VRAM allocation on gfx1151 (Strix Halo).
-# TheRock gfx110X-all packages provide compatible Python 3.11 wheels.
-# Install AFTER vLLM to override the CUDA torch it pulled in.
+# TheRock gfx110X-all packages provide compatible Python 3.12 wheels.
+# Reinstall AFTER vLLM to override the standard ROCm torch it pulled in.
+# vLLM's ROCm C-extensions remain compatible (same HIP ABI, torch 2.10.x).
 RUN --mount=type=cache,target=/home/ray/.cache/uv,uid=1000,gid=1000 \
     uv pip install --system --reinstall \
         --index-url https://rocm.nightlies.amd.com/v2/gfx110X-all/ \
@@ -101,9 +108,9 @@ RUN --mount=type=cache,target=/home/ray/.cache/uv,uid=1000,gid=1000 \
 # OOM or "insufficient GPU memory" at startup. The .pth file auto-patches
 # mem_get_info on Python startup to return sysfs VRAM values.
 COPY --chown=1000:100 amdsmi-shim/strixhalo_vram_fix.py \
-    /home/ray/anaconda3/lib/python3.11/site-packages/strixhalo_vram_fix.py
+    /home/ray/anaconda3/lib/python3.12/site-packages/strixhalo_vram_fix.py
 RUN echo "import strixhalo_vram_fix" > \
-    /home/ray/anaconda3/lib/python3.11/site-packages/strixhalo_vram_fix.pth
+    /home/ray/anaconda3/lib/python3.12/site-packages/strixhalo_vram_fix.pth
 
 # Pre-download common models for faster cold starts (optional, increases image size)
 # RUN python3 -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('BAAI/bge-large-en-v1.5')"